The price of AI is crashing faster than the rate of Moore's Law, report suggests — intelligence costs are in freefall, outpacing comparative technologies like compute, DNA sequencing, and lithium batteries | Tom's Hardware

A new report from AI research firm Epoch AI suggests the price of artificial intelligence has fallen by thousands of times in recent years — faster than any other transformational technology in the past century. The report suggests that the cost of artificial intelligence is just under 50% cheaper every quarter, and 13 times cheaper per year.

That's faster than lithium batteries, faster than DNA sequencing, and even faster than compute, which has been benefiting from the rapid advancements from Moore's Law for much of the past century

Although these falling costs are likely aiding adoption, it raises difficult questions about the necessity of remaining tied to frontier AI development, and could make it harder for third-party AI services to build customer loyalty and consistent revenue.

If you know that at any given time it's only a few weeks or months before the service you're interested in is going to be more capable and cheaper, what's the incentive to adopt it right now? If a competing service offers something better and cheaper, why stay with your current provider?

That poses a difficult conundrum for labs like OpenAI and Anthropic, which are racing towards ever greater AI capabilities. As the report suggests, finding large profits even with advanced models that outstrip the competition is difficult to imagine when it may only be a fleeting advantage.

• Frontier AI faces pricing reckoning as token volume explodes 25-fold
• Firm that lets you rent out your idle Mac for AI inference workloads warns of plummeting token costs
• Deepseek's latest AI model is smarter than ever, but its low running costs come at a price

Falling faster than DNA sequencing and 20th Century compute

Epoch's report paints a stark picture of the relative cost of what it terms "artificial thought." It shows the cost of AI falling by hundreds of thousands of times within just five years. Comparatively, it shows the price of Lithium Batteries between 1991 and 2024 as falling just 100 times over that near-25 year period. Compute fell the most over its lifetime, dropping 100s of billions of times over 60 years from 1940.

This is particularly relevant considering Moore's Law has remained a relevant constant throughout much of the past 60 years. With compute performance and efficiency improving by notable margins every year, the compound effect is modern computers that are both dramatically faster than their predecessors and vastly cheaper to run when measured against their compute capabilities.

AI performance and costs for that performance are falling even faster.

The metric that showcases the most similar fall in cost on the chart is DNA sequencing — an oft-cited example of the falling cost of technology. Its relative cost has fallen by an even greater extent than AI, but that took just over 20 years.

Epoch AI shows artificial intelligence doing the same in just half a decade. That works out to just under 50% cheaper every quarter, and 13 times cheaper per year. The fall is accelerating at a pace that's four times faster than DNA sequencing, and 18 times faster than lithium batteries.

What to read next

• China's open-weight AI models are now just 4 months behind frontier US offerings, Mozilla report claims
• TypeSafe AI's Jev offers an alternative to LLMs that claims to be 193x faster and 445x cheaper
• Anthropic’s IPO filing shows most of its compute goes to R&D, not serving customers

There are several caveats to these results. Compute only covers the period up until 2001, with no numbers on the relative cost of it after that, and electricity is only tracked until 1973, which means it completely misses the explosion of solar energy in recent years. AI cost falling was also only tracked from 2023 onwards, too, with data prior to that date extrapolated from the existing shorter trend and outside research.

Epoch AI also didn't base these results on a particular model, but on models achieving an 81.25% or better result on the Graduate-Level Google-Proof Q&A (GPQA) Diamond benchmark, and tracked the cost of having the model answer one of the multiple-choice questions on the test.

That does measure AI capabilities across a range of knowledge disciplines, but it doesn't necessarily cover the entire range of AI capabilities. It can show how much easier these kinds of tasks are for newer models, though, because they're not just scoring higher on the test more readily; they're doing it for far less, too.

How much cheaper?

Epoch AI cites OpenAI's GPT o3 model, released in January 2025, as achieving a 75% score on the GPQA Diamond test with a price of $0.30 per question. Just a year and a half later, though, OpenAI released GPT-5.6 Luna, which can score just as well on the same test, but costs only $0.0004 per question.

But that kind of rapid cost reduction isn't uniform across different disciplines, or linear in its progression. Expanding its research to include additional benchmark results on chess puzzles and mathematics, Epoch AI discovered vastly different rates of cost efficiency improvements. Where its GPQA Diamond benchmark showed periods of rapid cost reduction followed by periods of stagnation and plateauing, the AIME OTIS Mock test showed much more regular progression at each data point. Chess puzzles, although not tracked to the same success percentage, showcased more regular and consistent improvement, too.

Frontier Math developments, however, barely improved between 2025 and 2026, and then the costs fell off a cliff midway through the year.

Epoch AI also highlights that the rate of cost reduction is in decline, and that appears to be the case for most measured metrics for any particular performance level. Across the five benchmarks, it tracked cost falls of 66% per quarter initially, but two years later, that's down to 32% per quarter.

That could suggest AI progression is slowing, or that it's becoming harder to cut AI use costs, particularly in 2026, as the price of energy and compute hardware has skyrocketed. It's hard to offer a cheaper service if the raw materials for AI intelligence are vastly more expensive than they used to be.

Benchmaxxing is still an issue

Arguably the biggest potential problem for this data is, as Epoch AI highlights, the potential for "Benchmaxxing." That involves AI developers specifically training their models to do well on benchmarks that are otherwise designed to test more generalized capabilities.

Fooling benchmarks is something software companies have always done — Nvidia and ATI got caught doing it with 3DMark in 2003, and thousands of Intel test results were invalidated in 2024 for doing much the same thing with another benchmark.

Epoch AI's benchmarking used randomized elements to try to avoid models from recognizing they're being tested or developers specifically training them to be effective at third-party tests. But that's not all tests, and the ones without it showed greater rates of decline, suggesting there is some measure of benchmark optimization going on in the data.

That doesn't invalidate the results, but it does warrant taking them with an ounce of skepticism.

Not everyone takes advantage of the savings

The major concern for AI developers, especially those developing frontier models while spending hundreds of billions of dollars on infrastructure, is that these results suggest there's little point in paying for any kind of privilege. Even though the likes of Fable, Mythos, and the latest OpenAI models are still expensive to run compared to their contemporaries, token costs are falling all the time as the flagship developers compete on the Pareto frontier — the point where efficient cost and high intelligence meet.

But if other models can achieve 90% of the same intelligence at a fraction of the cost, and that cost is only likely to fall in the weeks and months to come, then is there any need to pay for the latest features and capabilities? Especially if those other models end up being open-weight, meaning third parties can compete to offer the most efficient and affordable version of that model.

But like anything else people subscribe to, it's not just about price, and it's not just about intelligence or capabilities either. Familiarity is important: With UI, with workflows, with the rest of what your organization is running. Trust is huge with any kind of ongoing commitment, especially financial. Can you trust that new third-party service offering a cheaper model than the one you've used for the past year? Maybe, but is it worth the risk?

Switching to a new model means confirming the veracity of those new benchmarks, and trusting that prices won't change dramatically in the future (they probably will), potentially invalidating your savings. You have to trust that the system you were using won't just catch up a week from now, and that the new system doesn't have any bugs or privacy concerns.

Changing AI tools isn't as straightforward as just chasing cost or intelligence. While prices might be falling dramatically, that doesn't necessarily mean a subscription model for AI is impossible. Just harder to justify.